reasoning

A LessWrong post exploring how the Qwen model can be taught to reason about the 'why' behind its outputs, examining alignment and reasoning behaviors in a long-form write-up.

A user reports that Qwen 3.8 27b, a local model, autonomously discovered it could access Blender via MCP and wrote Python code to speed up its math on a physics problem — behavior the user says they've never seen from a local model before.

A blog post introducing "Weight-Space Programming," where frontier LLM GPT-6 Astra natively reads and edits neural network weights—changing only 0.016% of Qwen3.5-0.8B's weights to boost MATH-500 from 7.9% to 48.1% and tripling IFEval, without gradients or training.

Apriel Voice 27B Preview 是一个开源语音-语言模型，基于 Gemma 4 26B-A4B 后训练检查点构建，通过两阶段训练将语音理解、推理与工具调用统一到共享推理骨干中，并在 Big Bench Audio、Speech-BFCL 和 τ-Voice 等基准上取得有竞争力的成绩。

The paper proposes CRoS, a CoT causality metric that rewards scrambling the order of reasoning steps without altering their content, avoiding confounding effects of traditional perturbation-based metrics. Evaluated with RLVR on Gemma-3-4B over Reasoning Gym, CRoS improves out-of-domain accuracy (46.8% vs 43.8%), CoT quality, and CoT faithfulness over existing causality and outcome rewards.

StrictPRMBench introduces a replayable benchmark for auditing how fixed-score PRM aggregation rules (product vs. geometric mean) affect best-of-N selection, showing product yields higher deletion-induced selection instability across 14 cells with reproducible reanalysis artifacts.

分析显示 Grok 4.20 在 PokeBench 的 Pokémon Red 基准中卡在房子附近长达数百回合，尽管其自身推理日志正确识别出陷入循环却仍无法脱困，引发关于视觉与规划能力瓶颈的讨论。

A user shares frustration with Qwen3.8-Flash-Next's verbose, meandering responses on general research questions when run locally via llama.cpp, questioning whether the model can be improved beyond coding tasks.

A practical, evidence-based guide to GPT-6 Astra's real-world use cases, mapping its performance across 12 Scale AI SEAL benchmark categories to project recommendations, strengths, limitations, and cases where alternative models like Claude Fable 5.1 are preferable.

This paper systematically analyzes when and where socio-demographic information benefits pluralistic offensiveness modeling, showing gains are clearer at group/population levels than individual ones, and that explicit reasoning about demographics outperforms direct conditioning.

ThinkJev is a training-free method that builds fast (System 1) and slow (System 2) answering paths from a single hybrid-thinking open model, reading calibrated option probabilities directly or after variable-length thought, so it only pays for thinking where it statistically guarantees a target error rate.

A practitioner compares Qwen3 27B running INT8 vs INT4 AWQ over two weeks of heavy agent logs, finding INT4 dramatically increases reasoning-loop rates (1.9 vs 4.7 loops/1k calls), cuts task success from 92% to 78%, and the gap widens at long context above 100k tokens. INT8 and even BF16 also loop, so quantization amplifies rather than causes the issue.

Kyutai Labs 重新审视了 Neural Thickets 方法，指出在输出 logits 上加入随机偏置、其余权重保持不变，即可匹配甚至超越对模型所有权重加噪并保留最佳副本的效果。

Carnegie Mellon paper shows that pretraining, fine-tuning, and RL can collapse into a single training rule (MaxRL as likelihood training on correct answers), with blended data runs outperforming staged training on reasoning and coding benchmarks while preserving general skills.

A community discussion asking for real-world user experiences with Nanbeige4.3-33B, a 33B parameter open model featuring a 512K context window and strong benchmark results in reasoning, coding, and agentic tasks, despite concerns about high KV cache memory usage.

A Google paper shows that Gemini AI agents are far less likely to wrongly undo correct work after a user's unfounded complaint when they first log a 'Thinking-in-Bets' checklist of verified claims and confidence levels, cutting erroneous rework from 12.2% to 0.02% in patch review.

NVIDIA releases AlpaJudge, a family of learned judge models for evaluating autonomous-driving reasoning, with reference-based and reference-free checkpoints available on Hugging Face under the OpenMDW-1.1 license.

This paper studies auxiliary trajectory preferences (confidence-based and entropy-based metrics) to improve small language model self-improvement and transfer, showing UID-Conj. gains of about 2 points over unmodified TTRL and uniform-weighted SFT on AIME tasks.

Researchers report feeding their hardest career research problems into a new AI model, which solved them one after another across many different fields.

Mistral 发布 Mistral Large 4（Le Chonk）：一个 1 万亿参数、49B 激活参数的原生多模态 Mixture-of-Experts 模型，具备 1M 上下文窗口、可配置推理力度、160+ 语言支持，并以 Mistral Ecosystem Open-Weight License 开源权重发布。